Evaluating Personality Traits in Large Language Models: Insights from Psychological Questionnaires

Evaluation and psychometric validity2025ACMApproved editorial review

Authors: Pranav Bhandari, Usman Naseem, Amitava Datta, Nicolas Fay, Mehwish Nasim

Keywords: Large Language Models, Personality, Big Five, Persona, Model Evaluation

Source: Open primary source (opens in a new tab)

5
Authors
24
Findings
49
Limitations
12
Evidence

Editorial summary

English

This five-page WWW Companion 2025 paper administers five reworded personality inventories to GPT-4, GPT-4o-mini, Llama-3-8B-Instruct, Llama-3.1-8B-Instruct, and Llama-3.2-3B-Instruct. GPT-4o rewrites the items; all-MiniLM-L6-v2 requires cosine similarity of at least 0.7 to the original, and the authors state that items below the threshold are reconstructed under human supervision. Items are randomized, sent in batches of ten, and repeated 100 times at temperature 0 for OpenAI and 0.01 for Llama. Table 2 releases only the mean by model, inventory, and dimension. No code, complete prompts, reworded forms, raw responses, or reproducible configuration was located. The protocol neither removes nor measures training-data contamination. A semantically close paraphrase may remain recognizable and may itself have been generated by a model familiar with the instrument; a similarity threshold measures textual resemblance, not presence in training data. It also does not validate preservation of scoring, factor structure, or psychometric meaning. The system prompt frames the model as a helpful assistant and asks it to rate itself, so responses may reflect alignment, format obedience, and stereotypes about a helpful assistant rather than latent traits. Instrument selection contains conceptual errors. HEXACO-100 measures six factors; the paper drops Honesty-Humility and forces Emotionality and the remaining domains into five Big Five labels. The 60-item measure called NEO-PI-R is actually IPIP-NEO-60, a public-domain representation of NEO PI-R. TIPI uses a 1–7 scale while the other four inventories are presented on 1–5 scales. Nevertheless, dominant dimension is computed by averaging all five raw means without normalization, giving TIPI a wider numerical range. A sensitivity check linearly converting TIPI to 1–5 happens to preserve all five dominant labels, but changes their means and does not resolve construct nonequivalence. The coefficient of variation also does not measure what the prose suggests. Although 100 runs are mentioned, Table 3 is reproduced, for the first three models, by taking a population standard deviation over only four questionnaire-level means per dimension, excluding TIPI. Thus GPT-4 Neuroticism at 20.49% or Llama 3.1 at 33.69% describes disagreement among instruments, not fluctuation across 100 responses. No run-level SDs or intervals are reported. The Llama 3.1 and 3.2 figures are not fully reproducible from Table 2; for Llama 3.2, the published Extraversion CV of 21.93% lies outside the 24.08–25.93% range compatible with the displayed rounding. The descriptive averages do show a pattern: under these prompts and items, raw Agreeableness, Conscientiousness, or Openness scores tend to be high and Neuroticism low. The declared dominant dimensions are Agreeableness for GPT-4, GPT-4o-mini, and Llama 3.2, Conscientiousness for Llama 3, and Openness for Llama 3.1. Without human norms, common calibration, uncertainty, statistical tests, or behavioral validation, these numbers do not establish personality profiles. Causal interpretations about fine-tuning and design objectives are not tested. The defensible conclusion is narrow: five aligned models produce different response patterns across inventories and families under a reworded self-report protocol. The work shows that instrument, response scale, and prompt matter; it does not establish psychological traits, psychometric reliability, removal of contamination, model-family personality, or behavioral prediction outside the questionnaire.

Español

Este trabajo de cinco páginas, publicado en WWW Companion 2025, administra cinco inventarios de personalidad reformulados a GPT-4, GPT-4o-mini, Llama-3-8B-Instruct, Llama-3.1-8B-Instruct y Llama-3.2-3B-Instruct. GPT-4o reescribe los ítems; all-MiniLM-L6-v2 exige similitud coseno de al menos 0,7 con el original y los autores indican supervisión humana para reconstruir ítems por debajo del umbral. Los ítems se randomizan, se envían en lotes de diez y se repiten 100 veces con temperatura 0 en OpenAI y 0,01 en Llama. La Tabla 2 publica únicamente la media por modelo, inventario y dimensión. No se localizó código, prompts completos, versiones reformuladas, respuestas crudas ni configuración reproducible. El protocolo no elimina ni mide contaminación de entrenamiento. Una paráfrasis semánticamente próxima puede seguir siendo reconocible y haber sido generada por un modelo que conoce el instrumento; el umbral de similitud evalúa parecido textual, no presencia en datos de entrenamiento. Tampoco valida que la reformulación conserve puntuación, estructura factorial o interpretación psicométrica. El system prompt presenta al modelo como un helpful assistant y le exige autoevaluarse, por lo que las respuestas pueden reflejar alineamiento, obediencia al formato y estereotipos sobre un asistente útil, no rasgos latentes. La selección de instrumentos contiene errores conceptuales. HEXACO-100 mide seis factores; el artículo omite Honestidad-Humildad y fuerza Emotionality/otros dominios a cinco etiquetas Big Five. La escala citada como NEO-PI-R de 60 ítems es realmente IPIP-NEO-60, una representación pública del NEO PI-R. TIPI usa 1–7 y los otros cuatro inventarios se presentan en 1–5. Aun así, la dimensión dominante se calcula promediando sin normalización las cinco medias, dando a TIPI un rango mayor. Una sensibilidad que transforma TIPI linealmente a 1–5 conserva por casualidad las cinco etiquetas dominantes, pero cambia sus medias y no resuelve la falta de equivalencia entre constructos. El coeficiente de variación tampoco mide lo que sugiere el texto. Aunque se mencionan 100 ejecuciones, la Tabla 3 se reproduce, para los tres primeros modelos, calculando desviación estándar poblacional sobre solo cuatro promedios de inventario por dimensión, excluyendo TIPI. Por tanto, Neuroticismo 20,49% en GPT-4 o 33,69% en Llama 3.1 describen desacuerdo entre instrumentos, no fluctuación entre 100 respuestas. No se publican SD o intervalos entre ejecuciones. Las cifras de Llama 3.1 y 3.2 no se reproducen enteramente desde la Tabla 2; en Llama 3.2, el CV publicado de Extraversion 21,93% queda fuera del rango 24,08–25,93% compatible con el redondeo mostrado. Los promedios descriptivos sí muestran un patrón: puntuaciones crudas altas en Agreeableness, Conscientiousness u Openness y bajas en Neuroticism bajo este prompt y estos ítems. La dimensión dominante declarada es Agreeableness para GPT-4, GPT-4o-mini y Llama 3.2, Conscientiousness para Llama 3 y Openness para Llama 3.1. Sin normas humanas, calibración común, incertidumbre, pruebas estadísticas o validación conductual, esos números no establecen perfiles de personalidad. Las interpretaciones causales sobre fine-tuning y objetivos de diseño tampoco se prueban. La conclusión defendible es limitada: cinco modelos alineados producen patrones de respuesta distintos entre cuestionarios y familias bajo un protocolo de autoinforme reformulado. El trabajo sirve para mostrar que instrumento, escala y prompt importan; no demuestra rasgos psicológicos, fiabilidad psicométrica, eliminación de contaminación, personalidad familiar del modelo ni que los perfiles predigan conducta fuera del cuestionario.

Research question

What Big Five scoring patterns do five LLMs produce when responding to reformulated versions of five inventories, and how much do the averages differ across dimensions, instruments, and model families?

Method

Descriptive study without human participants. GPT-4o reformulates items from BFI, HEXACO-100, TIPI, MINI-IPIP, and IPIP-NEO-60; all-MiniLM-L6-v2 calculates similarity with the original and a threshold of 0.7 is applied with declared human supervision. Five LLMs receive the items randomized in batches of ten through a system prompt for numerical self-report. 100 iterations per model are declared with minimum temperature. Likert scores are averaged per inventory and dimension; the CV is calculated over four inventory means per dimension, excluding TIPI, and dominance over the raw average of the five inventories.

Sample: The nominal unit is five models: two OpenAI and three Llama. The article declares 100 iterations per model and questionnaire, but does not report how many responses failed, whether repetitions were identical at minimum temperature, or execution-level statistics. Table 3 uses four inventory averages per dimension as observations for the CV, not the 100 iterations. There is no human sample, norms, judges, or behavioral evaluation.

Findings

  • The definitive version is a paper from WWW Companion 2025, DOI 10.1145/3701716.3715504, pages 868–872, license CC BY 4.0.
  • The substantive content coincides with arXiv:2502.05248v1.
  • The five pages were rendered and visually inspected.
  • GPT-4, GPT-4o-mini, and three Instruct checkpoints of Llama 3/3.1/3.2 are evaluated.
  • GPT-4o generates the paraphrases and all-MiniLM-L6-v2 applies cosine similarity with a threshold of 0.7.
  • The article declares 100 iterations, temperature 0 for OpenAI and 0.01 for Llama.
  • Items are randomized and grouped in batches of ten.
  • Table 2 publishes means per model, instrument, and dimension, without dispersion across executions.
  • Agreeableness, Conscientiousness, and Openness tend to have high raw means; Neuroticism tends to be low.
  • The published dominance is Agreeableness in GPT-4, GPT-4o-mini, and Llama 3.2; Conscientiousness in Llama 3; Openness in Llama 3.1.
  • HEXACO-100 has six official factors, but the analysis retains only five labels.
  • The cited 60-item instrument is IPIP-NEO-60, although Table 1 calls it NEO-PI-R.
  • TIPI uses 1–7 and the other inventories are shown in 1–5.
  • The dominance mean mixes those scales without normalizing.
  • Rescaling TIPI linearly to 1–5 does not change the dominant labels in a sensitivity analysis, but reduces the means.
  • The CV is calculated over four questionnaire means per dimension, not over the 100 executions.
  • The CVs of GPT-4, GPT-4o-mini, and Llama 3 are reproduced exactly with population standard deviation.
  • The maximum published CV is Neuroticism 33.69% in Llama 3.1.
  • With the rounded Table 2, the minimum CV of Llama 3.1 is Agreeableness 12.25%, not Openness 12.41% as Table 3 reports.
  • The CV of Extraversion of Llama 3.2 is recalculated as 24.98%, not 21.93%.
  • The published 21.93% falls outside the possible range under the rounding shown for those four averages.
  • No code, data, complete prompts, reformulated items, or reproduction artifacts were found.
  • The paper provides no statistical test, interval, or human comparison.
  • The observed pattern may reflect assistant alignment and prompt design as much as inventory content.

Limitations

  • Paraphrasing items does not demonstrate that the originals or their equivalents were not in the training data.
  • Cosine similarity measures semantic resemblance, not training contamination.
  • A minimum threshold of 0.7 deliberately retains high similarity with the original items.
  • GPT-4o may reproduce constructs or language from instruments it encountered during training.
  • There is no extraction test, canary, comparison with new items, or baseline without paraphrases.
  • The original/reformulated pairs and similarity scores are not published.
  • Human supervision is not described: number of reviewers, criterion, disagreement, or qualification.
  • It is not validated that the paraphrases preserve key, polarity, difficulty, or item content.
  • Internal reliability, test-retest, factor structure, convergence, or invariance of the reformulated forms are not evaluated.
  • The helpful assistant system prompt may induce socially desirable and prosocial responses.
  • Anthropomorphic self-report has no clear referent for a model without life or behavior of its own.
  • Responses may be semantic stereotypes about assistants rather than internal states.
  • There are no neutral prompt, third-person, observer, behavioral, or natural dialogue conditions.
  • Complete system/user prompts for the five instruments or TIPI adaptations are not published.
  • The protocol states at least three administrations and then 100 iterations without reconciling both numbers.
  • Temperature 0 or 0.01 does not guarantee independence or determinism across calls.
  • No seed, top-p, max tokens, API version, date, region, or backend policy are reported.
  • Snapshots of GPT-4, GPT-4o-mini, or GPT-4o are not identified.
  • Exact revisions, quantization, hardware, or software of Llama are not identified.
  • Format errors, retries, invalid responses, or exclusions are not documented.
  • The 100 responses may be pseudoreplications nearly identical under minimum temperature.
  • Standard deviations, intervals, or distributions across the 100 executions are not published.
  • The CV in Table 3 uses four inventory averages and therefore has n=4 per model-dimension.
  • The text presents that CV alongside n=100, which may lead to misinterpreting it as variability across executions.
  • There are no CV intervals or sensitivity analysis to a single inventory.
  • The Llama 3.2 row does not reproduce from Table 2 even considering the published rounding.
  • The Llama 3.1 row selects a minimum different from the one recalculated with the visible values.
  • No full-precision figures are provided to reconcile the discrepancies.
  • HEXACO measures six factors and is not simply another five-dimensional Big Five inventory.
  • Honesty-Humility is omitted, a defining dimension of HEXACO.
  • HEXACO Emotionality is not simply equivalent to Big Five Neuroticism.
  • The article confuses NEO PI-R with IPIP-NEO-60.
  • The instruments differ in content, number of items, construction, and properties; their raw means are not interchangeable.
  • TIPI uses 1–7 and the other inventories 1–5, but dominance averages the five raw values.
  • Scores are not standardized by range, norms, or measurement error.
  • The dimension with the highest mean is not a validated definition of dominant trait.
  • There are no human norms indicating whether a mean is high, low, or typical for each instrument.
  • Models are not compared with humans under the same reformulated items.
  • There are no significance tests, effect sizes, multiple correction, or preregistered hypotheses.
  • Five models do not support general inferences about families, size, or fine-tuning.
  • Within each family, size, generation, data, and alignment change simultaneously.
  • Attributions to fine-tuning objectives and design are speculative: there is no ablation or access to training.
  • Radar plots may suggest comparability even though each inventory has distinct properties.
  • Temporal stability, backend change, or sensitivity to order/batch is not evaluated.
  • Randomizing and grouping by ten may change context; that effect is not separated from item content.
  • There is no validation in behavioral tasks, preferences, dialogue, or real decisions.
  • Other languages, cultures, base models, system prompts, or configurations are not evaluated.
  • Privacy, harm, or inappropriate use of personality profiles are not discussed.
  • No code, dataset, preregistered protocol, or supplementary material was located.

What the study does not establish

  • It does not establish that LLMs possess psychological traits or a stable personality.
  • It does not eliminate or quantify training contamination.
  • It does not psychometrically validate the reformulated questionnaires for LLMs.
  • It does not demonstrate that HEXACO, Big Five, and IPIP measure equivalent dimensions in models.
  • It does not demonstrate variability across the 100 executions through the published CV.
  • It does not demonstrate that a higher raw mean is a comparable dominant dimension.
  • It does not demonstrate that GPT-4 is more agreeable or Llama 3.1 more open in a human sense.
  • It does not demonstrate causal effects of fine-tuning, size, or model family.
  • It does not demonstrate consistency outside the helpful assistant prompt.
  • It does not predict model behavior in natural interaction.
  • It does not offer fully reproducible figures for Llama 3.1 and 3.2.
  • It does not allow replication without items, outputs, snapshots, seeds, and code.

Traceability

Scope: Full text

Version: WWW Companion '25, Companion Proceedings of the ACM Web Conference 2025, pages 868–872, DOI 10.1145/3701716.3715504, CC BY 4.0; content checked against arXiv:2502.05248v1

Consulted source: https://doi.org/10.1145/3701716.3715504

Review: Codex complete bilingual full-text fidelity pass, definitive publisher-version reconciliation, all-page PDF visual inspection, table arithmetic reproduction, response-scale sensitivity analysis, instrument-identity verification against primary sources, contamination-claim audit, psychometric-validity and reproducibility assessment; summaries written from the complete published paper and recalculated tables rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4; exact snapshot, provider endpoint and run date not reported
  • GPT-4o-mini; exact snapshot and run date not reported
  • Llama-3-8B-Instruct; exact checkpoint revision and runtime not reported
  • Llama-3.1-8B-Instruct; exact checkpoint revision and runtime not reported
  • Llama-3.2-3B-Instruct; exact checkpoint revision and runtime not reported
  • GPT-4o as the questionnaire-rewording model; exact snapshot and prompt not reported
  • sentence-transformers/all-MiniLM-L6-v2 for cosine-similarity screening

Instruments and metrics

  • Big Five Inventory, 44 items, reported 1–5 response scale
  • HEXACO-PI-R 100-item form, reduced by the paper from six factors to five reported dimensions
  • Ten-Item Personality Inventory, 10 items, official 1–7 response scale
  • Mini-IPIP, 20 items, reported 1–5 response scale
  • IPIP-NEO-60, mislabeled NEO-PI-R in Table 1, 60 items and reported 1–5 scale
  • GPT-4o paraphrasing followed by all-MiniLM-L6-v2 cosine similarity threshold 0.7
  • Helpful-assistant numeric self-rating system prompt
  • Unweighted raw mean across inventories for dimensional dominance
  • Population coefficient of variation across four questionnaire averages per dimension

Data used

  • No human participant dataset
  • Five questionnaire response matrices generated by five LLMs across 100 declared iterations
  • Only rounded aggregate means in Table 2 are published
  • No released reformulated items, run-level outputs, order permutations, seeds, logs or scoring files

Evidence and location

  • Definitive publication, DOI, license, and pagination: WWW Companion '25 publisher version, page 868, DOI 10.1145/3701716.3715504; Macquarie and UWA publication records checked 15 July 2026
  • Prompt and set of five instruments: Published paper Figure 1 and Table 1, pages 869–870
  • Paraphrases, similarity, temperature, and randomization: Published paper Sections 2.1–2.2, page 870
  • Models, 100 iterations, and declared CV calculation: Published paper Sections 2.2–3, page 870
  • Means of the five models and instruments: Published paper Table 2, page 870
  • CV, dominance, and exclusion of TIPI from CV: Published paper Table 3 and Sections 3.1–3.3, pages 870–871
  • Arithmetic reproduction, rounding limits, and scale sensitivity: reports/verification/article-179-table-audit.json
  • HEXACO has six factors: Official HEXACO-PI-R inventory website checked 15 July 2026
  • Identity of IPIP-NEO-60: Maples-Keller et al. IPIP-NEO-60 paper cited as reference 8 and PubMed record checked 15 July 2026
  • Official 1–7 TIPI scale: Official Gosling TIPI instrument and scoring page checked 15 July 2026
  • Absence of code and artifacts: Exact-title, DOI, arXiv-ID and GitHub searches checked 15 July 2026
  • Complete visual inspection: All five pages of the definitive WWW Companion '25 PDF rendered and visually inspected on 15 July 2026